Why Neural Networks Fail to Predict Moscow Exchange Markets—and Where ML Actually Helps Traders

At a meeting on machine learning in trading in Moscow, a sobering thesis was presented: neural networks are almost useless as a tool for direct price forecasting. However, ML really helps with narrower tasks — determining market mode, filtering out bad trades, and reducing execution slippage.

The Illusion of Price Forecasting

The market is still home to a popular fantasy: just show a model a chart of Sber or an index, and it will output the exact price for tomorrow. In practice, this falls apart very quickly. Price is too noisy, the market constantly changes, and any pattern found in historical data often disappears the moment the strategy goes live with real money. In high-frequency trading, the situation is even tougher: nanoseconds matter, and a heavy model can simply kill the system with latency.

"ML does not provide control over the market and is not a magic button

for 'money'."

That's why experienced market participants view machine learning not as a generator of ready-made signals, but as a supporting layer above the base strategy. Even if the model guesses the direction of movement more often than random, that's not enough. Between a beautiful forecast and money lies the order queue, slippage, broker and exchange commissions. As a result, a formally strong model can show losses where backtesting looked almost perfect.

Where ML Works

The most sensible thought from the discussion is simple: machine learning brings value where the task is narrow and verifiable. Instead of trying to predict the market itself, the model helps break it down into components: what mode it's in, which signals to skip, and where costs will eat into potential profits. This approach is much closer to the real practice of a solo algo trader, who cannot afford hedge fund infrastructure.

• Market phase determination: trend, sideways, volatility spike
• Signal filtering: which base strategy trades to skip
• Execution optimization: how to enter a position with less slippage
• Order book microstructure analysis: pattern search in order flow

For a retail trader, this is especially important because they often combine all roles at once: finding data, cleaning it, engineering features, building the model, testing hypotheses, calculating risks, and then deploying the result to production themselves. In this configuration, ML is useful not as a showcase with loud promises, but as a tool for specific narrow improvements — for example, turning off a trend strategy in a tight sideways market or avoiding a trade with poor expected execution.

Why Strategies Die

Most problems start not in the model, but earlier — at the data and hypothesis testing stage. If a day is missing in the candles, futures are merged with errors, or features accidentally contain information from the future, the algorithm instantly finds a non-existent pattern. On historical data it looks like the Holy Grail, but in live trading it ends very quickly. Hence the main principle: garbage in, garbage out. Data quality and experimental honesty matter more here than choosing between trendy libraries and architectures.

From the discussion emerges a useful filter of five questions that should be run before writing code: is there a specific task, is there honest data, can the idea be tested accounting for commissions and delays, does the forecast translate into real action, and does it produce economic value. If there's no confident answer to even one question, ML is being deployed too early. Otherwise, you get the typical trap: there's a forecast, but no profit after all the costs.

What This Means

For the market, this is a good signal of sobriety. Neural networks do not replace a trading system and do not eliminate risk, but they do help make targeted improvements to processes where humans are already hitting the limits of data volume or reaction speed. Another important conclusion — the market lacks experience sharing: as long as individuals repeat the same mistakes separately, ML remains not a profit accelerator, but an accelerator of expensive experiments.